S P A N A M AND E N G S P A N :
MACHINETRANSLATION
AT THE PAN AMERICANHEALTHORGANIZATION
Muriel Vasconcellos and Marjorie Le6n 1
Pan American Health Organization
525 Twenty-third Street, N.W.
Washington, D.C. 20037
1
P R O J E C T H I S T O R Y AND C U R R E N T S T A T U S
1.1
OVERVIEW
Spanish-English machine translation (SPANAM) has been
operational at the Pan American Health Organization
(PAHO) since 1980. As of May 1984, the system's
services had been requested by 87 users under 572 job
orders, and the project's total output corresponded to
7,040 pages (1.76 million words) that had actually been
used in the service of PAHO's activities. The translation
program runs on an IBM mainframe computer (4341
DOS/VSE), which is used for many other purposes as
well. Texts are submitted and retrieved using the ordinary word-processing workstation (Wang OIS/140) as a
remote job-entry terminal. Production is in batch mode
only. The input texts come from the regular flow of
documentation in the Organization, and there are no
restrictions as to field of discourse or fype of syntax. A
trained full-time post-editor, working at the screen,
produces polished output of standard professional quality
at a rate between two and three times as fast as traditional translation (4,000-10,000 words a day versus
1,500-3,000 for human translation). The post-edited
output is ready for delivery to the user with no further
preparation required.
The SPANAM program is written in PL/I.
It is
executed on the mainframe at speeds as high as 700
words per minute in clock time (172,800 words an hour
in CPU time), and it runs with a size parameter of 215 K.
Its source and target dictionaries (60,150 and 57,315
entries, respectively, as of May 1984) are on permanently mounted disks and occupy about 9 MB each.
While SPANAM continues to build its reputation as a
work horse, at the same time development is well
advanced on a parallel system that translates from
English into Spanish, ENGSPAN. Also written in PL/I,
ENGSPAN uses essentially the same modular system
architecture that has been developed for SPANAM, but it
is conceived on the basis of up-to-date linguistic theory
leading to rule-based strategies for the parsing of syntactic and semantic information. The overall policy is to
regularly upgrade SPANAM as breakthroughs become
available in the more sophisticated ENGSPAN. In this
way it has been possible to maintain ongoing production
with SPANAM while its capabilities are gradually
enhanced and expanded. Because of this dynamic mode
of development, information about the theoretical status
of either SPANAM or ENGSPAN is necessarily shortlived.
1.2
EARLY HISTORY: 1976-1979
The Pan American Health Organization, with headquarters in Washington, D.C., is the specialized international
agency in the Americas that has responsibility for action
in the field of public health. It comes under the umbrellas of both the Inter-American System and the UN family, serving in the latter instance as Regional Office of the
World Health Organization. In addition to its headquarters staff of 546 in Washington, PAHO has a field staff of
652 that supports both the operational programs in its 10
Pan American centers, located in eight different countries, and its 30 representational offices, in 28 countries.
Business may be conducted in any of the four official
languages: Spanish, English, Portuguese, and French.
The translation demand is greatest into Spanish, which
over the years has corresponded .to more than half the
total workload, and, after that~ into English.
The
demand for Portuguese is considerably smaller, and there
is only an occasional requirement for French.
In 1975 the Organization's administrators undertook a
feasibility study and determined that MT might be a
means of reducing the expenditure for translation. There
was already a mainframe computer, then an IBM 360, at
the headquarters site, and the decision was made to
l Respectively, Chief, Terminology and Machine Translation Program,
PAHO, and Senior Computational Linguist, Machine Translation
Project, PAHO.
Copyright1985 by the Association for Computational Linguistics. Permission to copy without fee all or part of this material is granted provided that
the copies are not made for direct commercial advantage and the CL reference and this copyright notice are included on the first page. To copy
otherwise, or to republish, requires a fee a n d / o r specific permission.
0362-613X/85/020122-136503.00
122
ComputationalLinguistics, Volume 11, Numbers 2-3, April-September 1985
Muriel Vasconcellos and Marjorie Learn
develop an MT system that would run on this installation
on a time-sharing basis. Work was to focus on the Spanish-English and English-Spanish combinations.
The
effort was to be supported under the Organization's
regular budget.
The intention from the outset was that MT should
articulate with the routine flow of text in PAHO. Postediting was considered to be unavoidable, since the
system would have to deal with free syntax, with any
vocabulary normally used in the Organization, and, in
time, with a large range of subjects and different genres
of discourse. No serious thought was given to a mode of
operation that would require pre-editing.
Initial work was begun in 1976. A team of three parttime consultants worked for the Organization for two
years, and one of these consultants remained with the
project for a third year. In the beginning the approach
drew upon a number of the principles that had evolved at
Georgetown University in the late 1950s and early 1960s
in the course of work on the Russian-English system
known as GAT (Georgetown Automatic Translation,
described in Zarechnak 1979).
The first language combination to be addressed was
Spanish-English.
The consultants had opted for this
direction recognizing that results could be available earlier than if they had started with English as the source.
Parallel efforts were concentrated on the architecture
itself of the system and the extensive supporting software. The period 1976-1978 saw the mounting of this
architecture and the writing of a basic algorithm for the
translation of Spanish into English. At the end of three
years the Spanish-English algorithm was in place, as well
as eight other PL/I support programs that performed a
variety of related tasks. The Spanish source dictionary
had been built to a level of 48,000 entries (at that time
the verbs required full-form entries), with corresponding
English glosses in the separate target dictionary. Work
on the dictionaries was supported by mnemonic, userfriendly software developed in 1978-1979 to facilitate
the operations of updating, side-by-side printing, and
retrieval of individual records. A corpus of about 50,000
words had been translated from Spanish into English.
Efforts from English into Spanish had produced one page
of text.
H u m a n resources during the period 1976-1978
consisted of the three part-time consultants together with
PAHO's contribution in the form of dictionary manpower
(total of 24 staff-months in the three-year period) and,
starting in 1977, half-time participation of the staff
terminologist, who assumed the responsibility of coordination. A full-time computational linguist was recruited
and assigned to the project in 1979.
The year 1980 was a turning point for MT at PAHO.
Advances came together which made it possible to move
into a production mode. To begin with, the computational linguist took full charge of the system software,
replacing the consultants. Up to that time production
had not been feasible because there was no morphologiComputational Linguistics, Volume 11, Numbers 2-3, April-September 1985
SPANAM and ENGSPAN
cal analysis of verbs: the failure to find a high percentage of inflected verbs had meant that many sentences in
random text were barred from even the most rudimentary
analysis. Thus the first order of business was to develop
the needed morphological lookup.
At the same time, the operational problems of text
input, another major impediment to production, were
also resolved. An interface established between the IBM
mainframe and the Organization's word-processing facility (then a Wang System 30) enabled MT to take its place
in the text-processing chain and tap into a large body of
text that had been made machine-readable for other
purposes.
A conversion program was written which
handled the differences in representation of characters,
solved ambiguities of punctuation, and made certain decisions about the format. From the time this program was
installed, any Spanish text that had been keyed onto the
word processor, regardless of the purpose for which it
was entered, was available for machine translation.
1.3
OPERATIONAL PHASE: 1980-PRESENT
As production gained momentum, the MT staff was
increased by the assignment of a full-time post-editor and
by greater participation of the terminologist as head of
the project.
Over the next two years, the sources of machine-readable text for SPANAM increased at a steady pace. The
use of word processing at PAHO expanded, and, in addition, another mode of input became possible through
optical character recognition (OCR). Whereas word processing had previously been restricted to special services
provided by a typing pool, after the installation of wordprocessing hardware throughout the headquarters building (Wang OIS/140), all program units eventually came
to participate in the text-processing chain. Furthermore,
the optical character reader (a Compuscan Alphaword
II), previously used only for Telex transmission, was
interfaced with the word-processing system; this meant
that existing typewriters could also be used as input
devices, and therefore that texts could be prepared in the
field and machine-read in Washington.
With accelerated production, improvements to
SPANAM have followed in tandem. From the beginning
it has been the policy, and continues to be so today, that
the output from production serves not only to meet the
purpose for which it was requested but also to provide
feedback for further development of the algorithm and
dictionaries. As post-editing proceeds, note is made of
recurring problems at all levels. Capture of this information at the time of post-editing saves much work later on.
The messages written by the post-editor on the side-byside text serve as a basis both for updating the dictionaries and for making enhancements, as feasible, in the
algorithm.
In this way the Spanish source dictionary had grown to
a total of 60,120 entries as of May 1984. Of this total,
9 4 % were bases or stems and 6 % were full forms, all
with corresponding entries in the English target. Since
123
Muriel Vasconcellos and Marjorie LeOn
1981 the incidence of not-found words in random text
has been well under 1 % - limited usually to proper
names, scientific names, new acronyms, and nonce
formations. Through coordination with the terminology
side of the program, the glosses have been increasingly
tailored to the specific requirements of PAHO. In addition, microglossaries have been established for various
users, so that specialized translations can be elicited.
In its four years of operation, SPANAM has become
not only wiser but more efficient as well. The program's
speed of run time has increased from 160 words per
minute to over 700 wpm. Yet the algorithm, even though
it has sustained a major reorganization into modular
structure and regularly undergoes enhancement, remains
approximately the same size (2,085 statements as of May
1984).
Further details about the working environment of
SPANAM are given in sections 2 and 6.
1.4
SYSTEM DEVELOPMENT SPANAM/ENGSPAN:
1981-PRESENT
In early 1981 a long-range strategy was decided on for
the continued improvement of SPANAM and the development of a parallel system from English into Spanish.
Two consultants from Georgetown University, Professors
Ross Macdonald and Michael Zarechnak, undertook
separate evaluations of SPANAM at that time. Their
recommendations led to the adoption of a combined
working mode in which improvements were to be introduced in SPANAM according to a predetermined schedule
while at the same time development began on the other
system, ENGSPAN.
Recognizing that each language
combination imposed a different set of linguistic priorities, the consultants nevertheless emphasized that greatly
expanded parsing was needed in both cases, especially in
the analysis of English as a source language. Such parsing, in turn, called for revision of the dictionary record in
order to allow for a broader range of syntactic and
semantic coding. It was felt that the basic modular architecture of SPANAM, as well as the dictionary record in its
essential format, should also be used for ENGSPAN. A
common architecture for the two systems meant that they
could continue to share the same supporting software.
Thus, improvements could migrate readily from one
system to the other; it wouid be easy for them to crossfertilize.
Having adopted this approach to development, with
each side to benefit systematically from the work being
done on the other, the project addressed its attention in
1981 to the enhancements that had been recommended
for SPANAM. Then, as the SPANAM effort tapered off,
time was devoted increasingly to ENGSPAN. By the end
of 1982, the ENGSPAN program and dictionaries (about
40,000 source entries, most of them with acceptable
glosses in the Spanish target) were in place.
SPANAM and ENGSPAN
Translation from English into Spanish has special
importance for public health in the developing countries,
and this fact provided the incentive for seeking extrabudgetary support from the U.S. Agency for International
Development (AID). In August 1983, AID gave the
Organization a two-year grant for the accelerated development of ENGSPAN. 2 This funding has made it possible
to have a second computational linguist for the grant
period, as well as consultants and part-time dictionary
assistants who have undertaken specific tasks within the
approved plan of work.
With the added manpower, the project has made
significant progress on the English-Spanish algorithm.
Particular focus has been placed on the development of a
parser using an augmented transition network (ATN),
which as of April 1984 was integrated into the rest of the
ENGSPAN program. The dictionary record has been
modified, without any increase in its overall size, so that
it can now accommodate 211 fields, as compared with 82
in the 1980 version of SPANAM. Deep syntactic and
semantic coding has been introduced for dictionary
entries corresponding to a sizable proportion of the
experimental corpus of 50,000 running words.
2 APPLICATION ENVIRONMENT
2.1
PRE-EDITING POLICY
As indicated above, it has always been expected that the
output of SPANAM and ENGSPAN would have to be
post-edited. There was no application of MT at PAHO
for which a customized language would be feasible.
Since post-editing was inevitable, it was felt that a preediting step would be anti-economic: the advantages to
be gained would not be sufficient to offset the added cost
of a second human pass. Moreover, in order for pre-editing to be worthwhile, the process would have to draw on
a high degree of linguistic sophistication, and adequate
manpower for this purpose would be scarce: the pre-editor would need to be well prepared not only in translation skills but also in knowledge of the algorithm, or at
least a number of its capabilities and limitations.
Thus, pre-editing in the linguistic sense has been ruled
out for SPANAM and ENGSPAN. In theory, a document
can be sent for execution by SPANAM without being seen
by any human eyes. If the operator has keyed in the
original Spanish document using normal in-house typing
conventions, no adjustments whatsoever are required.
With inexperienced operators, however, and with texts
read automatically by the OCR, the precaution is taken to
check the format, particularly the line-spacing and page
width, since deviations from the standard at that level
can disrupt the work of the algorithm.
Production texts are run only once. Demonstrations
are always performed on random text.
2 Grant BPE-5542-G-SS-3048-00, awarded to the Pan American
Health Organization under letter dated 3 August 1983.
124
Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985
Muriel Vasconcellos and Marjorie Le6n
2.2
2.2.1
P O S T - E D I T I N G POLICY
GENERAL P O S I T I O N
The multifaceted approach to post-editing is an important feature of SPANAM. On one level, consideration is
given to the user's needs and capabilities and to the
purpose of the translation. At the same time, specific
linguistic strategies developed by the project are often
used in order to minimize the recasting of certain
unwieldy constructions which frequently recur - mainly
the result of verbs in sentence-initial position in Spanish.
Finally, there is a series of word-processing aids that help
to speed up the physical process of editing and to deal
with pragmatic decisions in the SPANAM output which
are not handled by syntactic rules.
The degree of post-editing is determined by:
1. the purpose of the translation,
2. the user's own resources for editing,
3. the time frame, and
4. structural linguistic considerations in the text itself.
A text may be needed for information only, for publication, or for a variety of uses between these two extremes.
If it is to be edited by the requesting office, only the most
glaring problems are dealt with by the post-editor. On
the other hand, if it is to be published without much
further review, the post-editor devotes careful attention
to the quality of the text. These factors are determined
in a conference with the user at the time the job is
submitted. As to time constraints, it may happen that the
work has to be delivered under considerable pressure:
information-only translations of 20-25 pages may have to
be delivered within a couple of hours, and once a
40-page proposal for funding was delivered in polished
form the same day it was requested. With translation for
publication, however, longer periods are negotiated.
Contrary to what might be deduced from the nature of
PAHO's mission, SPANAM is asked to cope with a wide
range of subject areas and types of text. There have
been: documents for meetings, international agreements,
technical and administrative reports, proposals for funding, summaries and protocols for international data
bases, journal articles and
abstracts,
published
proceedings of scientific meetings, training manuals,
letters, lists of equipment, material for newsletters - even
film scripts.
In an open, "try-anything" (Lawson 1982:5) system
such as SPANAM, with its highly varied applications,
experience has led to the conclusion that post-editing
requires a trained professional translator.
Whereas
Martin Kay (1982:74) suggests that the person who
interprets machine output "would not have to be a translator and could quite possibly be drawn from a much
larger segment of the labor pool", the SPANAM experience suggests that this conclusion would be valid only for
technical experts working on a text for information
purposes only. Even in such cases, the technicians at
PAHO are encouraged to request a more careful translation of passages that are of particular interest.
Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985
S P A N A M and ENGSPAN
Only an experienced translator will be aware of the
words whose variable meanings are dependent on extralinguistic context. For example proyecto in Spanish can
mean 'project', 'proposal', or 'draft', and the choice
depends on full knowledge of the situation to which the
text refers. Esperar can mean 'hope' or 'expect', and the
distinction is essential in English - sometimes even
crucial. Such ambiguities require the attention of a translator with training, experience, good knowledge of the
subject matter vocabulary in both languages, and a technical understanding of what is meant by the text. Only a
person with this combined background is in a position to
make the choices that will fully reflect the intention of
the original author. Another area in which the translator's role is important is in interpretation of the degree of
intensity associated with relative terms. For example,
trascendente in Spanish can have much less force than its
English cognate, and the entire tone of a message may be
over- or underdrawn, depending on the interpretation
given to a key term of this nature. Indeed, it has been
the experience of SPANAM that users, even technical
experts, can misinterpret the glosses appearing in the
machine output and assign an altogether incorrect meaning in the process of "correcting" the text. The role of
the experienced translator is not to be underestimated.
2.2.2
LINGUISTIC STRATEGIES
In addition to experience in the interpretation of nuances,
the post-editor needs a strong linguistic background in
order to master the particular strategies that have proved
to be effective in the processing of SPANAM output. For
the inexperienced post-editor, the most time-consuming
task is the recasting that is deemed to be "required"
when machine-translated constructions turn out to be
ungrammatical or intolerably awkward in the English
output. SPANAM has addressed this problem by developing a series of "quick-fix" post-editing expedients
(QFP) for dealing with the typical problems of Spanishto-English translation. For example, certain maneuvers
are suggested as being useful in the case of fronted verb
constructions in Spanish, which occur frequently and
present difficulties for the standard SVO pattern in
English. The purpose of the QFP is to minimize the
number of steps required in order to make the sentence
work. Since it was a V(S)O construction that triggered
the problem in the first place, any solution that avoids
reordering will necessarily depart from one-on-one
syntactic fit. In other words, in the example of the fronted verb, one might try to see if the opening phrase, which
will be a discourse adjunct, a cognitive adjunct (terms
from Halliday 1967), or the main verb itself, could be
nominalized so that it might serve as the subject of the
sentence in English.
Such an approach manages to
preserve in the theme position (Halliday 1967) the cognitive material which had been thematic in the source text,
usually with a parallel effect on the focus position as well
(Vasconcellos 1985b). For this reason, the result is often
quite satisfactory, even compared with a translation that
125
Muriel Vasconcellos and Marjorie Leon
is syntactically more "faithful" (see Section 6 below and
also Vasconcellos, in preparation)• The examples below
compare QFPs with solutions that were actually proposed
by translator-post-editors (traditional human translation
- THT).
In example (1), the semantic content of the fronted
verb is reworked into a noun phrase that can serve as the
subject of the sentence. Time is saved by leaving the
rheme of the sentence untouched; only a few characters,
highlighted inside the box, were changed. Moreover,
additional speed was gained by making changes from left
to right, in the same direction in which the text is being
reviewed.
(1)
Durante 1983 se inici6 ya la transformaci6n
paulatina de estos planteamientos en acciones.
MT:
During 1983 t ~ e [ was initiated already [ the
gradual transformation of these proposals into
actions.
THT: During 1983 these proposals already began to be
gradually transformed into actions. (62
keystrokes)
QFP: During 1983 [ progress began toward [ the
gradual transformation of these proposals into
actions. (27 keystrokes)
In example (2), on the other hand, the adjunct itself is
nominalized, again with a significant saving of time and
keystrokes.
(2)
En este estudio se buscarfi contestar dos preguntas
fundamentales:
MT:
~
this study [ it will be sought [ to
answer two fundamental questions:
THT: In this study answers to two fundamental questions will be sought: (53 keystrokes)
QFP: This study l, seeks | to answer two fundamental
questions: (14 keystrokes)
Use of the foregoing approach, wherever feasible,
adds up to substantial economy, with apparently little or
no deterioration in the quality of the translation (see
Section 6 below)• However, knowing when and how to
make such changes requires considerable skill. This is
one more reason why the post-editor should have a
strong background both in translation and, if possible, in
hngulsttcs as well•
It is always emphasized in SPANAM that editorial
changes should be kept to the minimum needed in order
to make the output intelligible and acceptable for its
intended purpose•
SPANAM and ENGSPAN
attention has also been given to speeding up the post-edit
by automating as many of the recurring operations as
possible.
The SEARCH-and-REPLACE function on the word
processor is heavily used in post-editing. In addition,
SPANAM has a set of special aids developed for the
purposes of MT. Besides a full set of possible word
switches ( l x l , l x 2 , 2 x l , 2 x 2 , l x 3 , 3 x l , 3 x 3 , etc.),
there are routines that deal with the character strings that
most often have to be changed in SPANAM output• For
example, only a single "glossary" keystroke is needed to
perform the following editorial operations:
SEAR CH-and-DELETE:
the, of, there, to, in order to
SEAR CH-and-REPLA CE:
from~of, for~of, for~by, in order to - / f o r -ing, a/the,
which~that, who~that, every~each, among~between, such
as~as, some of the~some
The inventory can be changed or expanded at will.
2.2.4
From the discussion above, it can be seen that speed in
post-editing is achieved by a combination of strategies.
Some of the points made may appear on the surface to be
almost trivial, but yet they can add up to a significant
difference• One example of an apparently trivial factor is
the method of positioning the cursor under the string to
be-modified. Delays at this point can add up to a surprising proportion of total time spent on post-editing, since
they will occur with every change that is made. Informal
experiments suggest that the most efficient approach for
positioning the cursor is to always use the SEARCH key.
The "mouse" and the light pencil appear to be less effective. The slowest method, unfortunately, seems to be the
one that is most often used, namely simple manual striking of the directional keys. Since people tend to rely on
the directional keys unless otherwise trained, this point is
emphasized with the post-editors who work on SPANAM.
The staff of the project are constantly on the lookout
for new ways of saving time. All tasks are streamlined as
much as possible• A series of programs have been developed on the word processor for automating the housekeeping support that has to be done apart from
post-editing, and recently some of this work was made
even more efficient by passing it on to the mainframe
computer.
Printing is kept to a minimum; finished
production is delivered to the user either on a diskette or
by a telephone call notifying the office that the job is
available on the system•
2.3
2.2.3
WORD-PROCESSING STRATEGIES
The SPANAM post-editors work directly on-screen.
Experience has shown that post-editing on hard copy,
with the changes entered by a "word-processing
operator", is not a highly efficient mode. Accordingly,
126
OTHER TIME-SAVERS
POST-EDITING VIS-a-VIS OTHER ASPECTS
OF THE SYSTEM
In the SPANAM environment there is a close link
between post-editing and the other aspects of the system.
The staff post-editor has been trained to update the
dictionaries, and currently almost all the dictionary work
Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985
Muriei Vasconcellos and Marjorie Le~n
on SPANAM is done by this person. Required changes to
the dictionaries are proposed at the time of post-editing.
Hence there is no need to go through the text a second
time. Also, glosses and other solutions seem to come to
mind most readily when the whole text is actually being
worked on. The post-editor, if adequately trained in
updating, is in the best position to see what dictionary
changes are necessary in order to deal with the specific
constructions that tend to recur in production translations.
The post-editor also alerts the computational linguist
to areas where the algorithm needs improvement.
2.4
INTEGRATION OF THE SYSTEM INTO A COMPLETE
TRANSLATION ENVIRONMENT
SPANAM/ENGSPAN, the terminology activity, and the
traditional human translation unit are in the process of
being merged into a single program of language services
at PAHO. While human and machine translation even
now coordinate the workload to a certain extent, as of
May 1984 there was not yet any centralized screening of
incoming jobs. It is expected that such a triage will make
it possible to maximize the effectiveness and efficiency of
the respective services.
Combination of the two activities will also make for a
more rational utilization of the manpower available at
any given time, with the staff being assigned to a variety
of different duties, depending on both needs and skills.
It is also a goal to reduce a given person's day in front of
the screen from eight hours to six through the rotation of
assignments.
In the area of management, the SPANAM/ENGSPAN
programs on the mainframe computer are also helping
with labor-intensive operations for which the human
translation service is responsible: it is already performing
automatic word counts, and spelling check systems are
being developed for both English and Spanish.
In terms of the linguistic work of the human translator,
SPANAM/ENGSPAN can help to lighten the load in a
number of ways. To begin with, technical and scientific
terms are retrieved in context, which means that MT is a
sort of very efficient lexical data base. With an ordinary
LDB, the translator has to go to the terminal (which is
not usually at his desk for his use alone), sign on, and
initiate a search. After he has performed all the mechanical steps, there is still the possibility that the term is not
in the data base at all, and his effort will have been wasted. When this happens repeatedly over time, a cumulative frustration builds up. With SPANAM/ENGSPAN, on
the other hand, not only does the translator know immediately what translation has been assigned to the term, he
is also told its degree of reliability and whether or not it is
in the WHOTERM data base. The status of the terms is
indicated by small superscript symbols which can be
requested at the time the text is sent for translation.
WHOTERM is also on the word-processing system. Its
general file contains:
a definition of each term in
English, translation equivalents in up to four languages
Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985
SPANAM and ENGSPAN
besides English, synonyms if there are any, a reliability
code for the primary term in each language, scope notes,
and a subject code. In addition to the general file, it has
files with: names of organizational entities, full equivalents for abbreviations, scientific names of pathogens,
generic names of drugs in three languages, and chemical
names of pesticides with trade names cross-referenced to
them (Ahlroth & Lowe 1983).
SPANAM also aids the translator with its system of
microglossaries for specialized subject areas (see Section
4.3 below). When a text is known to deal with a certain
subject, the translator can request a corresponding microglossary which will contain alternate glosses. One or
more of these microglossaries can be specified at the time
the job is submitted. The translator can also have a
mieroglossary of his own in which he can store special
terms he prefers to use.
It is also possible for SPANAM to provide alternate
choices in the output entry, such as project/
proposal~draft, hope~expect, time~weather, etc., although
this is not the regular policy. These alternatives can be
stored in a microglossary. In the output, the undesired
translation is eliminated by striking a single glossary key.
If the translator provides feedback in the form of
suggested or requested changes in the dictionaries, the
updating can be done immediately. Some of SPANAM's
users have developed the habit of providing regular feedback, and this means that their translations become
increasingly tailored to their specific requirements.
While there is no doubt that SPANAM/ENGSPAN
reach their maximum efficiency when post-edited
on-screen, at the same time studies are being done on
ways in which a translator can dictate his changes so that
they can be entered by a word-processing operator working from a tape.
The human translation service stands to benefit, also,
from the sophisticated facilities that have been developed
on the word processor for editing and housekeeping
support.
3
GENERAL TRANSLATION APPROACH
Since the bulk of the Organization's translation work
involves only Spanish and English, the machine translation system was developed specifically for this pair of
languages. No consideration was given to using the
interlingua approach. The broad range of subject areas
to be dealt with precluded the use of a knowledge-based
approach or one based on a representation of the meaning of the text. Although the systems are currently
language-specific, significant portions of the algorithm
could be adapted for use in a system involving
Portuguese or French, the other official languages of the
Organization. Because SPANAM and ENGSPAN were
developed separately, they reflect different theoretical
orientations and utilize different computational techniques. At the same time, they have many features in
common.
127
Muriel Vasconcellos and Marjorie Le6n
SPANAM was originally designed as a direct translation system. The translation is produced through a
series of operations which analyze the Spanish source
string, transform the surface structure to produce a
syntactic frame for the English target string, substitute
the English glosses indicated by the results of the analysis, insert a n d / o r delete certain grammatical morphemes,
and synthesize the required endings on the English
words. The principal stages involved in the translation
algorithm are: morphological analysis and single-word
lookup, gap analysis, multi-word unit lookup, homograph
resolution, subject identification, treatment of prepositions, object pronoun movement, verb string analysis,
subject insertion, do-insertion, noun phrase rearrangement, target lookup, target synthesis.
ENGSPAN is a lexical and syntactic transfer system
based on the slot-and-filler approach to language structure. It performs a separate analysis of the English
source string, applies transfer routines based on the
contrastive analysis of English and Spanish, and then
synthesizes the Spanish target string. The principal stages of this algorithm are: morphological analysis and
single-word lookup, gap analysis, substitution and analysis unit lookup, sentence-level parse, transfer unit lookup,
target lookup, syntactic transfer, and target synthesis.
The program includes backup modules for homograph
resolution, verb string analysis, and noun phrase analysis,
which are called in if the sentence-level parse is unsuccessful.
4
4.1
LINGUISTIC TECHNIQUES
MORPHOLOGICAL ANALYSIS
SPANAM's morphological lookup procedure makes it
possible to find most Spanish words in their stem forms.
The algorithm recognizes plural and feminine endings for
nouns, pronouns, determiners, quantifiers, and adjectives; person, number and tense endings for verbs; and
derivational endings such as -mente/-ly.
Bound clitic
pronouns are separated from verb forms, and any accent
mark related to the presence of the clitic is removed.
Another subroutine adds missing accent marks when the
source word is written with an initial capital or in all capital letters. The components of compounds formed with
hyphens or slashes are looked up as separate words. A
few prefixes are also removed from words without a
hyphen.
ENGSPAN's morphological analysis procedure, known
as LEMMA, is called if the full-form is not found in the
dictionary and the word consists of at least four alphabetic characters. This procedure checks for the presence
of a number of different endings, including -'s, -s; -s, -ly,
-ed, -ing, -er, -est, and -n't. Each time an ending is
removed, the new form of the word is looked up.
LEMMA uses morphological and spelling rules and short
lists of exceptions in order to determine when to remove
or add a final -e, when the word ends in a double consonant, etc. If a lemmatized form of the word is found in
128
SPANAM and ENGSPAN
the dictionary, its record is checked to make sure that its
part of speech corresponds with the ending which was
removed. If LEMMA exhausts all its possibilities, the
word is checked against a small list of prefixes (re-, non-,
un-, sub-, and pre-). If one of these prefixes can be
removed, another lookup is performed. If this final lookup is unsuccessful, a dummy record is created for the
word and a gap analysis routine is called. " N o t - f o u n d "
words are initially considered to be nouns and given the
possibility of also functioning as verbs and adjectives.
Information from both LEMMA and derivational suffixes
is used in order to confirm or reassign the main part of
speech, as well as to confirm, remove, or add possibilities
for ambiguities.
The lookup strategy used in both SPANAM and
ENGSPAN keeps down the size of the dictionary while
allowing a good deal of flexibility. The dictionary coder
has the option of entering a word in its full form, in one
or more of its inflected forms, or in its stem form. With
irregular forms and homographs, the full form must be
used. For example, in the Spanish source dictionary the
only entries for the word esperar are the stem esper and
the v e r b / n o u n homograph espera. The English source
dictionary contains an entry for expect and unexpected,
but not for expects, expected, expecting, or unexpectedly.
4.2
HOMOGRAPH RESOLUTION
SPANAM deals with homographs at several different
stages of the program. Ambiguities that can be resolved
by morphological clues or capitalization are handled by
the lookup procedure. Proper names are also identified
at this stage. One-character words are distinguished
from letters of the alphabet after the lookup has been
completed. The homograph resolution module handles
other types of homographs by examining the surrounding
context.
The possible parts of speech for a word are indicated
in the dictionary record in a series of bit fields which
include: verb, noun, adjective, pronoun, determiner,
numerative, preposition, modifier, adverb, conjunction,
auxiliary, and prefix. Any combination of two or more
bits may be coded. Other sequences of bit codes are
used to distinguish between different types of pronouns,
adverbs, and conjunctions: relative, interrogative, nominal, adverbial, connector, compound, and coordinate.
The use of multiple-word substitution units reduces
the number of lexical ambiguities which must be resolved
by the algorithm. Analysis units may also be used to
selectively specify the part of speech of any or all of the
words covered by the unit.
ENGSPAN's front-line approach to homograph resolution is embodied in the ATN parser, described in Section
5.3. The English words can be coded for the same possible parts of speech as in SPANAM. Determination of the
function of each word depends on the path taken through
the network. The sequence of parts of speech which
leads to the first successful parse is used as the basis for
the transfer stage.
Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985
Muriel Vasconcellos and Marjorie Leon
There are three ways in which lexical information
from the dictionary is used to help the parser arrive at the
correct analysis. Substi;ution units compress idioms into
one record with a single part of speech. Analysis units
can be used to indicate that a group of words can be
expected to occur in a collocation with a particular function. This information may be overridden, whenever
necessary, by the parser. An individual word may also be
coded to indicate which of its possible parts of speech is
statistically most frequent. Again, the final decision is
made by the parser based on the results of the sentencelevel analysis.
4.3
POLYSEMY
SPANAM/ENGSPAN have two principal tools for dealing
with polysemy:
microglossaries and transfer units.
Substitution units and analysis units are also used when
common collocations are involved.
A microglossary is a sub-dictionary of target glosses
which can be set up for a particular subject area,
discourse register, or specific user. Glosses pertaining to
the subject area of international public health form part
of the main dictionary. Microglossaries are in use for
special translations of terms in the fields of law, finance,
sanitary engineering, statistics, and scientific research.
The system may have up to 99 microglossaries with any
number of entries in each one. The microglossaries to be
consulted during the translation of a particular text are
specified at run time. The existence of a specific microglossary entry is indicated in the target record containing
the principal gloss for the word. Thus, no time is wasted
looking for special translations of every word. More than
one microglossary may be activated for the same translation, in which case they are listed and consulted in
order of priority.
The transfer unit is a rule that is stored in the source
dictionary and is retrieved after the analysis of the
sentence has been completed. The existence of a transfer unit is indicated in the record corresponding to the
individual source word. A transfer unit contains a condition to be tested and an action to be performed. Examples of conditions are:
• Subject of this verb has X feature(s) or is word W.
• Object of this verb (or preposition) has X feature(s)
or is word W.
• This word modifies a word with X feature(s) or modifies word W.
• This word has N object(s).
• Context N word(s) to left/right contains word with X
feature(s).
Transfer units are explicitly ordered in the dictionary.
The action may either select an alternative translation,
insert a word such as a preposition, or delete one or more
words. The action also indicates whether or not additional transfer entries should be sought for the same
word.
Computational Linguistics, Volume 1 I, Numbers 2-3, April-September 1985
S P A N A M and E N G S P A N
4.4
SYNTACTIC AND SEMANTIC FEATURES
The dictionary record for each lexical item (including
substitution units) contains bit fields that are used to
store information about its syntactic and semantic
features. These features are used in both the analysis
and transfer stages of the translation.
For example,
verbs and deverbal nouns are specified as occurring with
one or more of the following codes: no object, one
object, two objects, complement, no passive, locative,
marked infinitive, unmarked infinitive, declarative clause,
imperative clause, interrogative clause, gerund, adjunct,
bound preposition, and object followed by bound preposition. Subject and object preferences can be specified as
_+Human, +Animate, and +Concrete. Other fields are
reserved for case frames. Features which can be coded
for nouns include Count, Bulk, Concrete, Human,
Animate, Feminine, Proper, Collective, Device, Location,
Time, Quantity, Scale, Color, Nationality, Material,
Apposition, Body part, Condition, and Treatment.
Adjectives are coded for m a n y of the same features
mentioned above. In addition, they can be coded as
Inflectable, Optionally Inflectable, General, Temporary
condition, Positive connotation, and Negative connotation. Adverbs can be coded as Time, Place, Manner,
Motive, Interruptive, and Connector. One of the references used in developing the coding scheme for the
English entries was Naomi Sager's Natural Language
Information Processing ( 1981).
4.5
A N N O T A T E D SURFACE S T R U C T U R E N O D E S
In ENGSPAN, the structure produced by the parser
consists of a graph containing nodes corresponding to
each clause and phrase. Each node contains a list of its
constituents, their roles, and their locations.
If the
constituent is a lexical item, the location is a word
number; if it is a phrase or a clause, the location is the
pointer to the appropriate node. Each node is annotated
with features applicable to the type of phrase or clause
involved. These features include Type, Mood, Person,
Number, Tense, Aspect, and Voice.
Both the ATN formalism and the structural representation used in ENGSPAN draw heavily on the presentation of ATN parsers and systemic grammar in Winograd
(1983). Winograd's discussion, in turn, is based on the
work of Woods (1970, 1973) and Kaplan (1973). Of
course, the ATN parser has necessarily had to be adapted
to the needs and computational environment of the
PAHO project.
4.6
SPANISHVERB SYNTHESIS
The procedure for the synthesis of Spanish verb forms is
based on principles of generative morphology and
phonology. The program synthesizes all regular and
most of the irregular verbs, in all tenses and moods
except the future subjunctive, and in all persons except
the second person plural. The verb is entered in the
target dictionary in its stem form. Binary codes are used
129
Muriel Vasconcellos and Marjorie LeOn
to specify the conjugation class and the 11 exception
features which govern the synthesis of the irregular
forms. Only one dictionary entry is needed for each
verb. A small number of highly irregular stems and full
forms (74 in all) are listed in a table. The majority of
"stem-changing" verbs require no special synthesis
coding. The procedure consists of a series of morphological spellout rules; raising, lowering, diphthongization,
and deletion rules based on phonological processes;
stress assignment rules; and orthographic rules to handle
predictable spelling changes.
5
COMPUTATIONAL TECHNIQUES
5.1
DICTIONARIES
The SPANAM/ENGSPAN dictionaries are VSAM files
stored on a permanently mounted disk. The source and
target dictionaries are separate files. The basic record
has a fixed length of 160 bytes. The source entry is
linked to its target gloss by means of a 12-digit lexical
number (LEX). The first six digits of the LEX are the
unique identification number assigned to each pair when
it is added to the dictionary. The second half of the LEX
is used to specify alternative target glosses associated
with the same source entry. The main or default target
gloss for each pair has zeroes in these positions.
5.1.1
SINGLE-WORD ENTRIES
The key for a source entry is the lexical item itself, which
may be up to 30 characters in length.
The source
dictionary is arranged alphabetically. The key for a
target entry is the LEX, and the target dictionary is
arranged in numerical order.
Words may be entered in the source dictionary either
with or without inflectional endings. Most nouns are
entered only in the singular and adjectives only in the
masculine singular. Verbs are entered as stems. Fullform entries are required for auxiliary verbs, words with
highly irregular morphology, and homographs.
Several source items may be linked to the same target
gloss by assigning them to the same LEX. For example,
irregular forms of the same verb or alternative spellings
of a word require only one entry in the target dictionary.
Likewise, more than one target gloss can be linked to the
same source word through the lexical number. In this
case, each alternative gloss is distinguished by coding in
the second half of the LEX. Two positions are used to
designate terms belonging to microglossaries, two for
glosses corresponding to different parts of speech, and
two for context-sensitive glosses which are triggered by
transfer units.
5.1.2
MULTIPLE-WORD ENTRIES
The dictionaries contain four types of multiple-word
entries: substitution units (SU), analysis units (AU),
delayed substitution units (DSU), and transfer units (TU).
The key for a multiple-word entry in the source dictionary is a string consisting of the first six digits of the LEX
130
SPANAM and ENGSPAN
for each word in the unit. In the case of an SU or an AU,
the words must occur consecutively in the sentence in
order for the unit to be activated. A DSU or a TU can
cover either a continuous or discontinuous string.
The basic SU contains from two to five words. A
different record structure is used for longer entries, such
as names of organizations and titles of publications.
When an SU is retrieved, the dictionary records corresponding to the individual words are replaced with one
record corresponding to the entire sequence. The gloss
for the unit is also found in a single entry in the target
dictionary. This type of unit is essential in order to
obtain the correct translation of names of organizations,
titles of publications, slogans, etc., and is an efficient way
of handling some fixed idioms, phrasal prepositions, and
certain technical terminology. An SU record has the
same format as a single-word entry. In addition, it
contains a character string which indicates the part of
speech of each of its members. This information can be
used by the parser if it is unable to parse the sentence
using the single part of speech specified by the unit.
Examples of phrases entered as SUs are by leaps and
bounds, International Drinking Water Supply and Sanitation Decade, and Health for All by the Year 2000.
The AU, which also contains from two to five words,
has several functions. At the very least, it alerts the analysis routines to the possible presence of a common
phrase and provides information on its length and function. It can also be used to resolve the part-of-speech
ambiguity of any of its members. Finally, it can specify
an alternative translation for one or more of its parts.
The AU is an entry in the source dictionary but has no
counterpart in the target dictionary. The record for each
source word is retained in the representation of the
sentence, but the last two digits of its lexical number are
modified if a translation other than the main gloss is
desired. When the target lookup is performed, the gloss
for each word is retrieved separately. This ensures that
the rules for analysis and synthesis of conjoined modifiers will be able to access information about the individual
words of the phrase. It also makes it possible for the
parser to determine whether or not the individual words
are being used as a unit in the given context. Examples
of phrases entered as AUs are drinking water and patient
care. The algorithm is still able to correctly analyze
sequences such as the children have been drinking water
with a high fluoride content, and it is essential that the
patient care for himself.
The DSU is used to handle lexical items such as phrasal
verbs which are likely to occur as noncontiguous words
in the input. The existence of a DSU is indicated in the
source record of the first word of the unit. The unit is
retrieved from the dictionary during the sentence-level
parse. The decision of whether or not to accept the unit
is based on both syntactic and semantic requirements of
the parser. If the unit is accepted, it replaces the individual records and causes a different target gloss to be
Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985
Muriel Vasconcellos and Marjorie Le~n
retrieved. Examples of DSUs are look up, put on, and
carry out.
The TU is used to specify an alternate translation of a
word or words which depends on the occurrence of a
specific word or set of features in one of its arguments or
in a specified environment. These entries are stored in
the source dictionary only and are retrieved after the
analysis "has been completed. If the conditions specified
in the transfer entry are met, the corresponding lexical
numbers are modified so that the desired target gloss is
selected during the target lookup. For example, if the
object of know is coded a s + H u m a n , t h e verb is translated
as conocer instead of saber. If female and male modify a
noun coded as - H u m a n , they are translated as hembra
and macho instead of mujer and hombre.
5.2
GRAMMAR RULES
SPANAM, up to now, uses two basic types of grammar
rules: pattern matching and transformations. Pattern
matching is used for the recognition and reordering of
noun phrases. The grammatical patterns are stored in a
file which can be updated without recompiling the
program. The patterns are applied by searching for the
longest match first. Transformations are used to identify
and synthesize the verb phrases and clitic pronouns. The
rules are expressed in PL/I code and are grouped in
modules according to the part of speech of the head
word.
Each group of rules is tested once for each
sentence.
The structural description of each rule is
compared with the input string. The description may
require a match of parts of speech, syntactic features, or
specific lexical items. If a match is found, the rule is
applied. The rule may permute, add, delete, or substitute
lexical items or features associated with them.
As indicated earlier, ENGSPAN's grammar rules are
expressed in the form of an ATN. The network configuration indicates possible sequences of constituents. The
rules governing the acceptability of any specific input
string are contained in the conditions attached to the
various arcs of the network. The building of the nodes of
the structural representation and the assignment of
features and roles is determined by actions associated
with each arc. The conditions and actions are contained
in separate modules which are part of the compiled
program. The configuration of states and arcs is specified in a file which is updated on-line. The contents of
this file also determine which of the conditions and
actions are actually attached to specific arcs for a particular run.
As of May 1984 the ATN grammar had seven
networks: sentence, clause, noun phrase, verb phrase,
sentence nominalization, hyphenated compound, and
prepositional phrase. Each network consists of a set of
states connected by arcs. Four types of arcs are used:
category arcs, which can be taken if the part of speech
matches that of the input word; jump arcs, which can be
taken without matching a word of the input; seek arcs,
which initiate recursive calls to a network; and send arcs,
Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985
SPANAM and ENGSPAN
which return control to the calling network after the
successful parsing of a constituent.
5.3
PARSINGALGORITHM
The ENGSPAN algorithm performs a top-down, left-toright sequential parse using a combination of chronological and explicit backtracking. The parser stops after
completing the first successful parse. The path taken
through the network depends on the ordering of the arcs
at each state, the structural information already determined by the parser, and the codes contained in the
dictionary record for each lexical item and multiple-word
entry. Also available to the parser is information regarding sentence punctuation, capitalization, parenthetical
material, etc., which has been gathered by an earlier
procedure. The algorithm processes the words of the
input string one at a time, moving from left to right. At
each state, all arcs are tested to determine whether they
may be taken for the current word. The possible arcs are
placed on a pushdown stack and the top arc on the stack
is taken. The parser continues through the input string as
long as it can find an arc that it is allowed to take. If no
arc is found for the current word, the parser backtracks.
Which of the alternative arcs is taken off the stack
depends on the situation which caused the parser to
backtrack. If the end of the string is reached and the
algorithm is at a final state in the network, the parse is
successful. If no path can be found through the network,
the parse fails.
Long-distance dependencies such as those involved in
relative clauses and WH-questions are parsed by using a
hold Hst. When the parser encounters a noun phrase
followed by a relative pronoun, a copy of the phrase is
placed on the hold list. When a question is being parsed,
the questioned element is placed on the list. When a gap
is detected in the relative clause or interrogative
sentence, the phrase on the hold list is used to fill the
appropriate slot.
Whenever backtracking is required, a w e l l - f o r m e d
phrase list is used to save a copy of the phrases that have
been completed but are about to be modified or rejected.
For all seek arcs, the parser checks to see if a phrase of
the appropriate type is already on the well-formed phrase
list. If there are several phrases on the list that begin
with the same word, the longest phrase is tried first. A
new phrase is parsed only if there is nothing on the wellformed phrase list that satisfies the seek arc. In this way,
large amounts of reparsing are avoided.
Conjoining is currently being handled by a configuration of arcs at the end of each subnetwork which allows
additional phrases of the same type to be parsed recursively. When partial phrases are conjoined, the end of
the subnetwork is reached by traversing one or more
jump arcs.
In the event of an unsuccessful parse, ENGSPAN is
still expected to produce a translation.
The longest
successful path is always saved, and information from
this "partial parse" can be used by the synthesis routines.
131
Muriel VasconceHos and Marjorie Ledn
Local routines are used to analyze the remainder of the
input string. These routines function as a "safety net".
They resolve homograph ambiguities and analyze verb
strings and noun strings, adding as much information as
they can to the structural description of the sentence as a
whole.
The ATN parsing algorithm is being developed in an
independent PL/I program, using the ENGSPAN input
and dictionary lookup modules. 3 It is also totally compatible with SPANAM. The network grammar is read in at
runtime, making it possible to experiment with different
network configurations without recompiling the program.
Each time an enhanced version of the parser has been
tested and debugged, it replaces the working version in
the ENGSPAN program. The diagram in Figure 1 shows
the relationship between the parser and the "safety net"
routines in ENGSPAN. The parser is to be incorporated
in a similar way into SPANAM as well.
A complete description of the ATN grammar and
parser will be found in the report to be submitted to the
U.S. Agency for International Development at the end of
the grant period (October 1985).
6
6.1
6.1.1
PRACTICAL EXPERIENCE
SYSTEM MAINTENANCE
DICTIONARIES: SPANAM
As of May 1984, the Spanish source dictionary had
60,150 entries and the English target had 57,315. The
program for updating the SPANAM dictionaries is userfriendly. Many default codes are entered by the update
program automatically. Even though there are now 211
possible fields in which codes can be entered, as opposed
to the original 82, almost all of them can be specified
using mnemonic descriptors and code names.
Today, updating is done almost exclusively on the
basis of production text. Every job reveals ways in which
the dictionaries can be improved, with either new glosses
for individual words or idiomatic phrases, especially in
the case of technical terminology, or deeper coding of
existing entries. The steady, ongoing development of the
dictionaries (Table 1) has ensured both a decrease in
not-found words, with advantages for program effectiveness, and closer correspondence to the type of language
used in the Organization, leaving less work for the posteditor.
As indicated earlier, it is the post-editor who notes the
changes needed at the time of post-editing and who later
updates the dictionaries. An hour is reserved for this
work at the end of the day. When production permits,
the post-editor may spend extra time on dictionary work;
S P A N A M and ENGSPAN
if there is pressure, the work may have to be postponed
for a while. Because of the integration of dictionarybuilding into the work of the post-editor, the cost is no
longer an element that can be clearly identified.
T a b l e 1.
Size of dictionaries, PAHO Machine
Translation System, 1976-1984.
SPANAM
ENGSPAN
Year
Spanish
English
English
Spanish
1976
4,000
3,500
1977
7,836
7,341
1978
38 506
38 376
1979
48 289
53 303
1980
5O 912
55 792
1981
53 785
51 187 1
44,411
1982
54 383
52 223
40,107
41,358
1983
56 247
53 326 3
40,772
42,116
May 1984
60 150
57 315
41,210
42,638
2
44,998
7,000 unmatched target entries were deleted by a special-purpose
program.
2 Upon reversal of dictionaries, 4,500 duplicate source entries and
corresponding target records were deleted by a special-purpose
program after selection of the desired gloss.
3 1,000 irregular verb forms were deleted by a special-purpose program.
6.1.2
DICTIONARIES: E N G S P A N
ENGSPAN has the same user-friendly software as
SPANAM for updating its dictionaries. As of May 1984,
the English source dictionary had 41,210 entries and the
Spanish target had 42,638.
The AID project has provision for two half-time
dictionary assistants, one a linguist of English mother
tongue and one a translator of Spanish mother tongue. A
new deeply coded source entry costs from $0.60 to
$1.00; Spanish target glosses that require research are
about the same. Simple changes in existing entries average about $0.25 each.
6.1.3
O T H E R SYSTEM M A I N T E N A N C E
The SPANAM programs are maintained by the staff
computational linguist, who carries out these tasks in
addition to development work on ENGSPAN.
Hardware and system support are provided by the Pan
American Health Organization. There is no separate
charge for utilization of the computer, either for time or
storage space. The project has a permanently assigned
partition of 512 K in core, as well as 1 MB on disk for
work space, 8.5 MB for program libraries, and 33.2 MB
for dictionaries and other permanently mounted files.
3 A major portion of the parsing routines have been developed by Lee
A n n Schwartz, who has participated in this activity on a full-time basis
since August 1983.
132
Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985
Muriel Vasconcellosand Marjorie Ledn
SPANAM and ENGSPAN
PROCEDURES
GRAMMAR AND LEXICON
MANGIqTS
DATA STRUCTURES
Jy
[
INPUTTEXT IN
INTERNAL FORI'LJI,
T
wrsIoW
INPUT STRING
'DICTIONARY RECORDS
FOR SINGLE WORDS
LOOKUP
ENGLISH SOURCE
DICTIONARY
1-)
FINDUNIT
I
~
'I
"ODIFIED RECORDS
FOR WORDSt PHRASES
SENTENCEREGISTER
PARSER
CLAUSE REGISTERS
ATN GRN4HAR
VERBSTRING
PHRASEREGISTERS
POSN,I~ I 6
NOUNSTRING
TRANSFERUNIT
LOOKUP
SPANISH TARGET
DICTIONARY
,1
TLOOUP
MODIFIED CODES
FOR WORDSt PHRASES
J"
.,4"1--- SPANISHAND
CODEsGLOSSES
TSYNTAX
SPANISH
OUTPUTSTRING
HTSI Ot~
Figure 1.
I
[
FOR~TTED
OUTPUTTEXT
The relationship b e t w e e n the parser a n d the "safety n e t " routines in ENGSPAN.
ComputationalLinguistics, Volume 11, Numbers 2-3, April-September 1985
133
Muriel Vasconcellos and Marjorie Le6n
6.2
6.2.1
TESTING
SPANAM
The output of SPANAM is subject to daily scrutiny by the
project staff. In addition, weekly demonstrations are
given for visitors, using random text as input.
An experimental version of the program is used to test
and debug system enhancements resulting from
production feedback and from the research and development being done for ENGSPAN. Before the experimental
version of the program replaces the production version, a
control test is performed by translating the same text
with both programs and comparing the output, using the
Document Compare software available on the Wang
OIS/140. This utility program compares the output character by character. The CPU time, throughput time, and
number of disk I / O ' s for both versions are also
compared.
6.2.2
ENGSPAN
An experimental corpus of over 50,000 words was
selected at the beginning of the project. Sentences are
chosen from this corpus for the testing of specific
program modules. Following every major enhancement
of the algorithm or dictionary, the corpus texts are
retranslated and the results are compared with previous
translations. The system is also tested using new texts
which are translated without previous review by the
project staff. After the first translation run, the dictionary is updated in order to add not-found words and missing codes, and then the translation is rerun. Problem
sentences from these random texts are retained for use in
subsequent development tasks.
6.3
6.3.1
S P A N A M and ENGSPAN
M E A S U R E M E N T CRITERIA
SPEED: C P U / T H R O U G H P U T TIME
SPANAM's speeds, both CPU and throughput time, have
steadily improved over the years (Table 2). The best
throughput time speeds are obtained at night, when there
are fewer users working on the computer. During the
day, turnaround at peak periods can be considerably
slower. The speed is adequate for the current load of
production.
When major changes are made in the program, care is
always taken to make sure that they do not cause any
extensive degradation of speed.
The CPU and throughput times for ENGSPAN are
1,400 and 400 words per minute, respectively. These
speeds are expected to decrease as the coverage of the
ATN grammar is expanded.
Table 2.
Translation speeds, SPANAM, 1979-1984.
Best clock time
Year
wpm
1979
Average CPU time
wph p a g e s / h
wpm
wph
160
9,600
38
Notavailable
1980
176
10,560
42
Notavailable
1981
192
11,520
46
3,184
191,000
1982
580"
34,800
139
2,600
156,000
1983
700
42,000
168
2,880
172,800
1984
710
42,600
170
2,982
178,920
*Reflects change to VSAM hookup.
6.3.2
C O M P A R I S O N W I T H H U M A N T R A N S L A T I O N SPEED
With a trained post-editor, SPANAM's output is never
slower than that of a human translator. The range is
from one and a half to four times as fast, with the average falling between two and three times as fast. The
SPANAM output ranges from 4,000 to 10,000 words a
day per post-editor, depending on all the factors
mentioned above, as well as the sheer difficulty of the
text. On the other hand, human translators working in
the international organizations commonly produce
around 2,000 words a day, although some services report
an average of 1,500 and others an average of 2,500. It is
possible to reach 3,000 or even higher, but usually not on
a regular basis. Free lances report much higher rates.
Given the variability of both sets of figures, it would be
difficult to make any hard-and-fast comparisons.
However, for the same person using both modes, it might
be possible to draw some conclusions: one translator
who post-edits for SPANAM reports that she consistently
produces about three times as much output with MT as
she does in the traditional way.
6.3.3
QUALITY: C O R R E C T N E S S
No systematic error analysis has ever been done of
SPANAM or ENGSPAN. Three consultants were engaged
under different contracts to evaluate the overall status of
the project: Professors Yorick Wilks (1978), Ross
Macdonald (1981), and Michael Zarechnak (1981).
While they commented on general characteristics of the
output, they were more concerned with underlying processes that might produce the errors than the errors as
such. Referring to the quality of the output, Professor
Macdonald (1981:7) reported:
The current output is rather good. If a human being had
written it, perhaps the output would be considered to be
defective in many respects. When it is known, however,
that it was produced by a machine, the basis of judgment
shifts, and the output seems really very presentable. Any
person of good will can understand this output, and I
134
Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985
Muriel Vasconcelios and Marjorie Le~n
assume that no misleading translations have been discovered that would vitiate the intent of any article.
6.3.4 QUALITY:POST-EDITING EFFORT/HUMAN TRANSLATION QUALITY
The question of effort required for post-editing is inextricably tied up with standards of human translation. Both
these issues are highly colored by subjective criteria. In
Section 2.2 there was a discussion of linguistic strategies
for reducing the time spent on post-editing. The "quickfix" post-edit takes much less time than a traditional
revision.
In an effort to see how translators would handle some
of the same sentences that had been fixed up quickly in
post-editing, a set of 17 Spanish source sentences was
given to 12 trained translators who were asked to provide
spontaneous human versions in English (Vasconcellos, in
preparation). No one sentence was translated twice in
the same way; apart from lexical differences, there was a
variety of combinations and permutations in the ordering
of the various phrasal elements. However, when the
respondents were subsequently shown the "quick-fix"
alternatives, they agreed that the latter were at least as
good, or in some cases even better, than what they themselves had proposed - probably because there was greater cohesion in the presentation of the semantic
components (Vasconcellos 1985b, in preparation). This
exercise underscored the difficulty of measuring the
quality of a translation.
6.4
COST EFFECTIVENESS
Because of all the variables involved, including in particular the purpose of the translation, it is usually rather
difficult to make clear-cut comparisons between
SPANAM production and traditional translation at PAHO.
On one large project, however, such a comparison was
possible because about half the original text was in Spanish and the other half was in English. The former was
done on SPANAM and the latter was farmed out to
human translators who worked in the traditional mode.
For 101,296 words of machine translation, the cost was
$3,218, including a hypothetical cost for machine time,
and 36 staff-days were devoted to the activity. Had the
same number of words been farmed out in the traditional
mode, the cost would have been $8,196 and the number
of staff- and contract-days (based on an output of 2,000
words a day) would have amounted to 65.75. Hence
there was a monetary saving of $5,078 ( 6 1 % ) and the
staff-days were reduced by 29.5 ( 4 5 % ) .
It 'is safe to say that the economy effeeted with
SPANAM is sufficient to cover the salary of the post-editor and perhaps another salary at the same level as well.
It may yet be some time, however, before the early
investment in the project is fully recovered.
Sometimes it is hard to know whether or not SPANAM
is translating text that would otherwise have been
submitted for human translation. Quite possibly the user
is less hesitant to request a machine translation than a
Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985
SPANAM and ENGSPAN
human one.
B. Dostert (1979) has reported this
phenomenon in a survey of 58 users of MT.
In the past PAHO has farmed out a large percentage of
its translation load. As SPANAM and ENGSPAN increasingly reduce that direct cost, there is clear evidence of
savings.
6.5
SUBJECTIVE FACTORS
The SPANAM staff have come to the understanding that
in the end an MT system will stand or fall depending on
the human environment in which it is placed, and that
some of the most important factors cannot be measured.
In the broad sense, these include: long-term commitment,
positive attitudes, innovative responses, creative problem-solving. At the more specific level, they include also
the real availability of input in machine-readable form, a
cooperative spirit among the staff who must share the
oversaturated word-processing equipment, willingness on
the part of the post-editor to use the word processor for
long periods, resourceful post-editing, and a host of other
factors of nonresistance that are seldom taken into
account.
In addition, if human translators are to be enlisted as
post-editors, they must have a positive attitude toward
the capabilities of MT, and, for true gains in productivity,
they must be willing to use the keyboard and to become
adept with the special editing features that have been
developed for the word processor.
In dealing with the output, there must be flexibility
and compromise in regard to quality. For example, if the
rapporteur of a meeting has an hour in which to write up
what her speakers said, and she can't understand the
Spanish without a translation, there must be a " c a n - d o "
type of staff that will produce a document that can be
worked from. The need met is the true criterion.
7 DISCUSSION OF APPROACH ADOPTED
Macdonald (1979:130-145) has pointed out that MT
systems tend to polarize toward either an empirical or a
theoretical approach.
Development of the empirical
system proceeds " o n the basis of actual experience with
appropriate texts" (1981:1), whereas the theoretical
approach begins by postulating the adequacy of some
particular model of language description which it is
hoped will be able to cover all contingencies (1979:130,
1981:1).
Each position has its advantages and its disadvantages
(1981:1). In an empirical approach, the main advantage
is that the system is more compact in that it concentrates
on only that which is of immediate usefulness for the task
at hand. The main disadvantage is that, as the system
expands, "it becomes difficult to add in new operations
without reworking some of what has already been done"
(1981:1). The theoretical system, on the other hand, is
able to proceed with less disruption, but the disadvantage
is that "it is extremely difficult to predict as to which of
all the complexities will actually arise; complexities may
135
Muriel Vasconcellos and Marjorie Leon
be foreseen and planned for in the system which do not
actually appear in the type of texts to be translated"
(1981:1).
Rather than advocate one extreme or the other,
Macdonald believed that benefit was to be gained from
both; ideally, he felt, the two positions should be
combined in a melded system (1979:143):
On the whole, the best approach is a compromise
between the two extremes, a basically empirical approach
in which, however, the researchers strive for an overall
perspective (1981:1) . . .
The preliminary research on the empirical system will
serve the purpose of establishing the nature and extent of
the problem of machine translation clearly and definitively. When the nature of the problem has been recognized
as fully as possible, a rigorous and elegant solution of the
problem can be devised (1979:145).
This view, which came through strongly in the 1981
recommendations of Macdonald and Zarechnak, is what
has guided the development of SPANAM/ENGSPAN in
the last three years. While SPANAM began as a purely
empirical system, ENGSPAN is now well on its way to
being a melded system. The flexibility of the basic
SPANAM/ENGSPAN architecture has made it possible to
introduce a theoretical focus while still preserving that
part of the working system which could be used both in
the interim and as a "safety net" in the event of failed
parses.
Rather than weighing the techniques of one MT
system against those of another, it is felt that a more
fruitful approach is to establish certain "positive criteria
of relevance":
• Who are the system's users?
• In what environment is it being implemented?
• What purpose does it serve?
• On what basis is its use to be justified?
The strength of a system will lie in its capacity to be relevant for its users, its environment, its purpose, its justification. Judgment from this standpoint is believed to be
more effective in the long run than evaluation of the relative success or failure of a given theory.
By these standards, SPANAM has proved itself through
its four years of ongoing production: through the accessibility of its programs, the user-friendliness of its dictionaries, its rapid throughput time, the broad range of text
for which it produces a usable translation, and the
savings it has effected in terms of both time and money.
ENGSPAN is soon to follow suit.
ACKNOWLEDGMENTS
The authors wish to express special acknowledgment to
their co-worker Lee Ann Schwartz.
As the second
computational linguist working on ENGSPAN, Lee has
given unstinting support to the project and its objectives.
Growth of the dictionaries is attributable largely to the
contributions of Rebecca Naidis, post-editor of SPANAM,
136
SPANAM and ENGSPAN
and Susana Santangelo, who has worked on ENGSPAN.
Other assistance on the dictionaries has been given by
Catherine Howe, David Landeck, Jos6 Luiz Meurer,
Patricia Schmid, and Lucretia Vanderwende.
The occasional consultants to the project since 1981
have been Ross Macdonald, Michael Zarechnak, Leonard
Shaefer, and William Cressey.
Within PAHO, the project has had continuing support
and encouragement over the years from Luis Larrea
Alba, Jr., Chief of the Department of General Services.
Finally, both PAHO and the project are most appreciative of the support received from the U.S. Agency for
International Development.
REFERENCES
Ahlroth, E. and Armstrong-Lowe, D. 1983 The WHO Terminology
Information System: Interim Report. HBl/ISS/83.I (offset), World
Health Organization, Geneva.
Dostert, Bozena 1979 Users' Evaluation of Machine Translation:
Georgetown MT System.
Undated typewritten report later
expanded into article of the same title in Henisz-Dostert et al.:
149-244.
Halliday, M.A.K. 1967 Notes on Transitivity and Theme in English.
Part 2. Journal o f Linguistics 3:199-244.
Halliday, M.A.K. 1979 Explorations in the Function of Language.
Elsevier North-Holland, Amsterdam, New York, Oxford.
Henisz-Dostert, B.; Macdonald, R. Ross; and Zarechnak, Michael
1979 Machine Translation. Mouton, The Hague.
Kaplan, Ronald M. 1973 A General Syntactic Parser. In: Rustin,
Randall: 193-241.
Kay, Martin 1982 Machine Translation. ,4JCL 8(2): 74-77.
Lawson, Veronica 1982 Machine Translation and People. Practical
Experience of Machine Translation. North Holland, Amsterdam,
New York, Oxford: 3-9.
Macdonald, R. Ross 1979 The Problem of Machine Translation. In:
Henisz-Dostert et al.: 89-145.
Macdonald, R. Ross 1981 A Study of the PAHO Machine Translation
System. Report of short-term consultant under Personal Services
Contract APO-09435(WU1) [typescript, 13 p.]. Pan American
Health Organization, Washington, D.C.
Rustin, Randall, Ed. 1971 Natural Language Processing. Algorithmics
Press, New York.
Sager, Naomi 1 9 8 1 Natural Language Information Processing: ,4
Computer Grammar of English and its ,4pplications. Addison-Wesley,
Reading, Massachusetts.
Vasconcellos, Muriel 1984 Machine Translation at the Pan American
Health Organization: A Review of Highlights and Insights. Newsletter of the British Computer Society, Natural Language Translations
Specialist Group 14.
Vasconcellos, Muriel 1985a Management of the Machine Translation
Enviornment: Interaction of Functions at the Pan American Health
Organization. In: Lawson, Veronica, Ed., Translating and the
Computer 5: Tools for the Trade. Aslib, London: 115-129.
Vasconcellos, Muriel
1985b Theme and Focus: Cross-Language
Comparison via Translations from Extended Discourse. Ph.D.
dissertation, Georgetown University, Washington, D.C.
Vasconcellos, Muriel In preparation Functional Considerations in the
Postediting of Machine-Translated Output. (To appear in Computers and Translation.)
Winograd, Terry 1983 Language as a Cognitive Process. Volume 1.
Syntax. Addison-Wesley, Reading, Massachusetts.
Woods, W.A. 1969 Augmented Transition Networks for Natural
Language Analysis. Report No. CS-I to the National Science Foundation, Cambridge, Massachusetts.
Woods, W.A. 1971 An Experimental Parsing System for Transition
Network Grammars. In: Rustin, Randall: 111-154.
Zarechnak, Miachel 1979 The History of Machine Translation. In:
Henisz-Dostert et al.: 1-87 [in particular, 29-30, 32, 134 ff.l.
Computational Linguistics, Volume I I, Numbers 2-3, April-September 1985